Hyperspectral Image Classification Based On Deep Convolution And Adversarial Network With Small Samples

Hyperspectral image classification is an important means to obtain information of ground objects in the field of remote sensing.In view of the characteristic of union of imagery and spectrum,how to extract the joint spatial-spectral features is a key problem in the task of hyperspectral image classification.In addition,the problem of small samples and the phenomenon of different objects which have the same spectrum in hyperspectral images also pose challenges to hyperspectral image classification methods.To solve these problems,we propose hyperspectral image classification algorithms based on deep convolution and adversarial networks.The main research contents are as follows:A convolutional neural network(CNN)algorithm based on spatial attention mechanism is proposed.This algorithm aims at capturing the spatial information which is beneficial to classification from the spatial window input of hyperspectral image centered on the current pixel,so as to improve the classification effect of hyperspectral image.Firstly,each pixel in the hyperspectral image is weighted by using the space mask learnt from the original hyperspectral image by CNN,so that the model pays more attention to pixels which are relevant to the central pixel and advantageous to classification,weakens the influence of the pixels which are not related to the central pixel or even disadvantageous to classification,and extracts the important part of the spatial neighborhood input of the hyperspectral image adaptively.Then the weighted output by adaptive spatial mask is used as the input of CNN to improve the classification effect of CNN for hyperspectral images.Experiments show that the proposed CNN algorithm achieves consistent improvement on different hyperspectral image datasets.At the same time,compared with other advanced algorithms,the advantages of the proposed algorithm are validated.A semi-supervised discriminant convolution neural network classification algorithm is proposed.This method aims to solve the problems of different objects which have the same spectrum,big differences within the same class and small samples in hyperspectral images.It extracts the deep features of hyperspectral images based on convolution network,and reduces the intra-class differences of the features used for classification.Specifically,for labeled samples,the distances between each feature and the feature center of the category which it belongs to is penalized;for unlabeled samples,the weighted distance between each feature and the feature centers of all categories is constrained.They can promote the network to extract more discriminatory features,improve the generalization performance and classification effect of the network.Furthermore,the use of unlabeled samples makes up for the shortage of labeled samples in hyperspectral images,reduces the risk of over-fitting,and effectively improves the classification accuracy.The robustness of the proposed method on different hyperspectral data sets and its advancement compared with other algorithms are validated by experimental analysis.Semi-supervised learning can alleviate the problem of small samples in hyperspectral images,but there are many unlabeled samples.How to select high-quality samples from many unlabeled samples is a problem.In recent years,generation antagonism network has attracted wide attention due to its powerful sample generation ability.From the perspective of sample generation,the problem of small samples in hyperspectral images is alleviated.We propose a hyperspectral image classification algorithm based on multi-class spatial-spectral generative adversarial network(GAN).The algorithm aims to extract spatial-spectral features and achieve hyperspectral image end-to-end classification by using multi-class spatial-spectral GAN.We construct two generators by adopting one-dimensional transposed convolution neural network and two-dimensional transposed convolution neural network,respectively,to generate spectral and spatial information of hyperspectral images.Then we use a discriminator based on one-dimensional CNN and two-dimensional CNN to extract spectral and spatial features and finally achieve classification.At the same time,we define a new objective function for multi-classification tasks.The purpose of the discriminator is to predict the real training samples as their classes,and to predict the generated samples as belonging to all classes with the same probability.The goal of the generator is to make the discriminator distinguish the generated samples as real classes and make mistakes.Through the adversarial learning between the generator and the discriminator,and with the help of generating samples,the discriminant performance of the discriminator is gradually improved.The experimental results on three hyperspectral image datasets show that the proposed algorithm is superior to other advanced classification algorithms.Especially in the case of fewer samples,this advantage is more significant.